Machine Learning image training on EC2 with GPU

I am training deep learning models with ten thousand images on a G4 GPU instance, using local storage. Using parallel PyTorch dataloaders, just like I do with on-prem GPU hardware. On-prem, GPU utilization is typically a constant 99% during training and varies during validation steps. On EC2, training flips between 30/maybe up to 70% util and back to zero, for an average of maybe 30-40%. Please suggest how to get more GPU utilization in this scenario.